Deep Semantic Segmentation Assisted Region-of-Interest Sensitive Deep Spatio-Textural Feature Learning Framework for Leprosy Detection and Classification

The last few decades have witnessed exponential rise in leprosy also called Hansen diseases globally. Being chronic and infectious in nature, eradicating leprosy has remained a challenge. Despite a few recent efforts employing vision computing-based leprosy detection and classification, most of the at hand solution remain confined due to their inability to address data-imbalance, feature optimality, reliability and scalability. Addressing aforesaid challenges requires a vision-based method to accommodate automated region-of-interest (ROI)-specific information-rich learning and classification. In this paper, a highly robust deep semantic segmentation assisted region-of-interest sensitive deep spatio-textural feature learning framework is proposed for leprosy detection and classification. The proposed method at first applies the different pre-processing methods including contrast adaptive histogram equalization, intensity equalization, resizing. In sync with probable data-imbalance problem, we applied deep semantic segmentation model to obtain the ROI specific regions, which were subsequently processed for RGB conversion. Thus, the obtained ROI-specific RGB images were processed for feature extraction by using deep spatio-textural feature extraction model encompassing GLCM and ResNet101 deep network. Here, GLCM extracted sufficiently large spatio-textural features, while the use of ResNet101 not only provided deep residual features but also alleviated the likelihood of gradient fading problem, which is highly frequent in the state-of-arts like convolutional neural network (CNNs) or other recurrent neural networks (RNNs). Thus, the obtained deep spatio-textual features processed for horizontal concatenation-based feature fusion, which was later trained by using random forest ensemble classifier.The simulation results revealed that the proposed model exhibits accuracy and F-Measure of 97.20% and 97.80%, respectively. Thus, the ability to yield superior accuracy along with the potential to alleviate class-imbalance, feature inferiority and gradient fading makes proposed model more reliable and efficient towards real-time leprosy detection and classification.